Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix issue with loras, lowvram and --fast fp8.

comfyanonymous3bb4dec720bd
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
index 6ca124e6d..2ba306335 100644
--- a/comfy/model_patcher.py
+++ b/comfy/model_patcher.py
@@ -88,8 +88,12 @@ class LowVramPatch:
self.key = key
self.patches = patches
def __call__(self, weight):
- return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+ intermediate_dtype = weight.dtype
+ if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
+ intermediate_dtype = torch.float32
+ return comfy.float.stochastic_rounding(comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype), weight.dtype, seed=string_to_seed(self.key))
+ return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
class ModelPatcher:
def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
self.size = size